Method for monitoring an electric machine

Abstract

The invention relates to a method for monitoring an electric machine, wherein the monitoring is carried out for recurring operating states of the dynamoelectric machine, wherein real data in relation to control of the dynamoelectric machine are monitored against expected data, wherein deviations between the real data and the expected data are added over time, wherein the deviations are used to classify the behaviour of the control, wherein the classification of the control is monitored over time, wherein the expected data are changed on the basis of the monitoring of the classification.

Description

[0001]202214353 Foreign version 1 Description Method for monitoring an electrical machine The invention relates to a method for monitoring an electrical machine. An electrical machine is, for example, a dynamoelectric machine, such as a motor or a generator. An electrical machine is, for example, a drive which has a power converter and a dynamoelectric machine, such as in particular a motor. The power converter itself can also be referred to as an electrical machine. Electrically driven machines such as pumps, fans or compressors can also be referred to as electrical machines. The invention relates in particular to a method for the automated detection of speed control deviations in drive converters, which are also to be subsumed under power converters, for standard applications such as pumps, fans or compressors.The electric machine is controlled primarily by means of a closed-loop control system. There is a deterministic relationship between the parameters of the target speed and the reference motor speed. In many cases, the two parameters normally agree. However, additional parameters, such as a minimum speed or the avoidance of critical speeds, influence the relationship between the parameters. This means that there is no generally applicable rule, and machine-specific differences arise. For example, when the torque limits are reached, the target speed may not be reached. Such conditions, events, or developments must be recognized and monitored. The target speeds and actual speeds of the electric machine can be monitored manually by experts over longer periods. An analysis of a time series, e.g.at irregular time intervals, serves in particular to classify the current health status of the electrical machine. In this way, the electrical machine and / or the application of the electrical machine can be monitored. 202214353 Foreign Version 2 One object of the invention is to improve the monitoring of an electrical machine. One solution to the problem arises from a method according to claim 1. Another solution to the problem arises, for example, from one of claims 2 to 7.In a method for monitoring an electrical machine, wherein the monitoring is carried out in particular for recurring operating states of the electrical machine, speed-related data is used for monitoring, wherein a linear regression model is learned based on training data, wherein a total error sum is calculated, wherein if a limit value is exceeded by the total error sum, an error is detected and / or a classification takes place and / or a new model is learned. In particular, a classification takes place. The classification has two states, such as the normal state and the abnormal state. If the classification always has the abnormal state over a predetermined or predeterminable time, a new model is learned. If, for example, the determined total error sum exceeds the limit value, abnormal behavior is detected. The system is therefore classified as abnormal.If the limit is not exceeded, the system is classified as normal. If the classification is abnormal for a longer period of time, a new model is trained. A new model is therefore trained when the error or when the total error sum is permanently present, i.e. the total error sum is permanently exceeded. The duration for which a permanent exceedance occurs is, in particular, adjustable. The new model is, in particular, based on a linear regression model. The limit used for the total error sum is, in particular, adjustable. The duration can, for example, be seconds, minutes, or hours and can, in particular, be set by a user. The set or used duration depends in particular on the application. The limit can, in particular, also be adjusted by a user.For example, the method can be adapted for specific machines. In particular, abnormal operating states of machines can be detected using a general monitoring approach. By using a total error sum and monitoring it using a limit value, it is possible to continually adjust the monitoring of a machine, i.e. to adapt it repeatedly, especially if framework conditions should change. This can result in an endless loop of recurring adaptations for monitoring. By using the linear regression model in conjunction with the formation of a total error sum, reliable automated detection of algorithm malfunctions can be achieved. In one embodiment of the method, the error used for summation is set to zero for irrelevant points in time.In one embodiment of the method, a weighting is used to calculate the total error sum. This weighting specifically relates to a forgetting factor. The forgetting factor weights errors according to their age. The older an error is, the less weight it receives. This prevents past errors from being carried over into the future.In a method for monitoring an electrical machine, monitoring is carried out for recurring operating states of the dynamoelectric machine. Real data from a control system of the dynamoelectric machine is monitored with expected data. Deviations between the real data and the expected data are accumulated over time. The deviations are used to classify the behavior of the control system. The classification of the control system is monitored over time. The expected data is changed depending on the monitoring of the classification. 202214353 Foreign version 4 Recurring operating states include, for example, the starting or braking of a motor. A recurring operating state can also be an operating state of a system or depend on this operating state of the system, whereby the electrical machine is located in the system.The data used for monitoring is primarily speed data, which can also be referred to as parameters. For example, in a motor, there is a deterministic relationship between the parameters of the target speed and the reference motor speed. In many cases, the two parameters normally agree. However, additional parameters, such as a minimum speed or the avoidance of critical speeds, influence the relationship between the parameters. This means there is no generally applicable rule, and machine-specific differences arise. In particular, when the torque limits are reached, it can happen that the target speed is not reached. Until now, the target and actual speeds had to be monitored manually by experts over long periods of time. An observation of the time series at irregular intervals served to classify the current state of health.However, this evaluation requires expert knowledge, since in many operating states, a deviation between the signals is irrelevant. This is due, for example, to motor start-up and stop-down, load changes, process-related speed setpoint changes, etc. Inverter parameter settings, such as a minimum speed or the avoidance of critical speeds, are also influencing factors that affect the relative motor speed. Currently, these parameters are read manually and used by experts for evaluation. This type of problem can be solved using the method described. This is achieved in particular through an automated comparison of the parameters and a classification into normal or abnormal behavior.Automatic data preprocessing and a data-driven approach enable the determination of the expected normal state and can be used to automatically identify anomalies within a short period of time. A change (e.g., process parameter) is also automatically detected based on the data. The error status is adjusted accordingly. This enables a continuous comparison, in particular of the target / actual motor speed, with automated consideration of the operating state and machine-specific dependencies. An anomaly detection system based on this completes the overall system for monitoring the key parameters. Speed-related data, in particular, is used as data for monitoring. Expected data is, in particular, data generated by a model. Using the model, at least part of the electrical machine can be modeled.The electrical machine can be simulated, i.e. modeled, at least partially or entirely. In particular, a control system for the electrical machine is modeled so that the behavior of the electrical machine can be fully or partially simulated. In particular, actual values ​​of the electrical machine are simulated. These are, for example, a speed, a torque, a temperature and / or values ​​(data) dependent thereon, etc. Values ​​arise in particular from signals from the electrical machine. Signals are generated in particular by sensors. In this way, an electrical machine can be monitored. This makes it possible to automate monitoring, in particular of a speed-controlled electrical machine. Monitoring is carried out in particular for or in non-dynamic applications of the electrical machine.An example of the use of the described monitoring in a non-dynamic application is an application in which a target speed or an actual speed of the electrical machine is constant. The monitoring method is carried out in particular taking into account at least one of the following boundary conditions or criteria: 202214353 Foreign version 6 ^ Detection of process-related speed changes and differentiation from incorrect behavior (adaptation of the error detection to normal process-related changes), ^ Detection of control system-related speed changes and differentiation from incorrect behavior and / or ^ Detection of parameter-related speed changes and differentiation from incorrect behavior. This can be achieved by systematic adaptation of the detection, without active knowledge of the above-mentioned changes.These criteria can otherwise only be filtered in complex manual data aggregations in order to detect faulty control. The automatic dynamic adaptation of the detection algorithm to changed boundary conditions without adjusting parameters or long learning phases, taking into account a generic approach (as described here) and with the help of a numerical model of the electrical machine, in particular the motor, the converter and / or the load machine, is an example of a basis for the described fault detection. The monitoring of the electrical machine can also concern or include an evaluation of the electrical machine. Such an evaluation can replace expert knowledge. Operating states can be taken into account. In some operating states, a deviation of the values ​​orSignals (data) from one another, in particular the deviation of an actual value from a setpoint, is not relevant. This is caused, for example, by motor starting and stopping, load changes, process-related changes to the speed setpoint, etc. Inverter parameter settings, such as a minimum speed or the avoidance of critical speeds, are also factors that influence the reference motor speed. These parameters can either be read out manually and used by an expert for evaluation, or they can be automatically integrated into the monitoring system. Another option is the setting of static limit values, which can trigger an alarm for the "in torque limit" state. This allows the state of the torque limit to be monitored, and a comparison of the parameters can also be carried out automatically. This eliminates the need for manual visual inspection.For example, specific expert knowledge is no longer necessary, which would require time-consuming, in-depth data aggregation. Analysis requires a high level of personnel effort. Automated monitoring, even over very long periods of electrical machine operation, means that manual observation is no longer required. During operation of the electrical machine, data is monitored over longer periods of time, with particular consideration given to exclusion criteria or events. Exclusion criteria or events arise in particular from operating states and are automatically taken into account. In one embodiment of the method, the data is or relates to a speed and / or torque of the electrical machine. In one embodiment of the method, the accumulation over time relates to minutes, hours, days and / or months.This results in longer periods of data monitoring. These periods differ from short periods, such as the times that occur during control. Controllers in electrical machines often operate in short periods such as seconds or milliseconds. In one embodiment of the method, a model relating to the electrical machine is used to generate the expected data. The model can, for example, also relate to the control of the electrical machine and to the electrical machine in whole or in part. 202214353 Foreign version 8 In one embodiment of the method, a new model is created depending on the classification. During or through the classification, it is therefore recognized, for example, whether the model is still correct. If it is still correct, it is used again. If it is no longer correct, a new model is created. This is done in particular by retraining the model, orCreating a new model and training it. In one embodiment of the method, the model and / or the new model are trained. This means that after training, the model or the new model is trained. In this way, a model can be adapted to a specific environmental situation for the electric machine. At least one of the described methods enables, in particular, a long-term comparison of a requested and obtained engine speed. In particular, an automated comparison is carried out, e.g. of the key parameters, and a classification into normal or abnormal behavior. Automatic data preprocessing and a data-driven approach enable the expected normal state to be determined and can be used to automatically identify anomalies within a short period of time. Changes in boundary conditions (e.g. process parameters) can also be automatically detected based on the data.An error status, for example, is adjusted accordingly. The method enables a continuous comparison of, for example, the target / actual engine speed, with automated consideration of the operating state and / or machine-specific dependencies. The anomaly detection based on this completes the overall system for monitoring the parameters. In one embodiment of the method, automatic data preparation is carried out. The automated data processing serves in particular to increase accuracy. It reduces the number of incorrectly detected anomalies. Differing parameter values ​​do not always indicate real problems, but arise due to measurement inaccuracies or insignificant operating states: ^ Smoothing signal: The target / actual signals are smoothed using a rolling median: ^^^ ^^^ ൌ ^^ ^^ ^^ ^^ ^^ ^^^^ ^^. ௧ିଶ , ^^ ௧ି^ , ^^௧ , ^^ ௧ା^ , ^^ ௧ାଶ ^^ ^ Consider the sign: The sign of the referenced motor speed is adapted to that of the required speed: ^^ ^^ ^^ ^^ ^^ ^^ ^^ ൌ ^^ ^^ ^^ ^^ ^^ ^^ ∗ ^^ ^^ ^^ ^^^ ^^ ^^ ^^ ^^ ^^ ^^^ ^ Determine relevance: The next step is to determine the relevance of a timestamp ^^. Measured values ​​that are a certain amount of time away from the next off state are considered irrelevant. Periods in which the machine is not running (including, for example, start-up and shut-down times) are not taken into account. Times in which the referenced speed is below the minimum limit specified in the converter are also considered irrelevant: These steps are carried out automatically so that an expert can concentrate directly on the essential events and is not misled by false alarms. The inclusion of the additional measured values ​​takes the context relating to the operating state and a required speed into account and enables an automated comparison of the relevant points in time. In one embodiment of the method, an automatic training phase is carried out. In the automatic training phase, the collected machine data is used to learn a reference model which describes the normal relationship between the requested (SpdDmd) and referenced (SpdRef) engine speed. 202214353 Foreign version 10 ^ To determine the relationship between the parameters, the data ^^ of the respective machine is used.The irrelevant measuring points marked in step 1 are filtered and ignored: ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ൌ ^ ^^ | ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^^ ^ A linear regression model is learned from the training data, which describes the relationship: ^^ ^^ ^^ ^^ ^^ ^^' ൌ ^^ ∗ ^^ ^^ ^^ ^^ ^^ ^^ ^ ^^ In one embodiment of the method, automated detection and classification are carried out. This occurs in particular in an application phase of the monitoring. In the application, the model learned in the above step is applied to new data and a total error sum is calculated. The reference model supplies the normally expected related engine speed for a requested speed.This is compared with the actual value, which leads to the error: ^^ ^^ℎ ^^ ^^ ^^^ ^^^ ൌ | ^^ ^^ ^^ ^^ ^^ ^^^ ^^^– ^^ ^^ ^^ ^^ ^^ ^^'^ ^^^| ^ For irrelevant points in time from step 1, the error is set to 0: ^^ ^^ℎ ^^ ^^ ^^^ ^^^ ൌ 0 ,∀ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^ To classify a timestamp as normal or abnormal, the error is added to a total error sum ∑. Larger deviations have a greater impact, as the error sum grows more quickly with larger deviations. The total error sum decreases due to a forgetting factor ( ^^ <1). over time and past errors are not permanently carried forward: ∑^ ^^^ ൌ ^^ ∗ ∑^ ^^ െ 1^ ^ ^^ ^^ℎ ^^ ^^ ^^^ ^^^ ^^^ ^ If the error sum exceeds an adjustable limit, the time is classified as abnormal: ^ The error sum is limited to a maximum ^^, whereby a correction is detected promptly. The forgetting factor reduces the error sum during normal operation, and after a sufficient number of normal measured values, the status is automatically set to normal. ∑^ ^^^ ൌ ^^ ^^ ^^^^∑^ ^^^, ^^ ^^ Using one of the described procedures, for example, an automated comparison of requested and received engine speed is possible without any manual effort. This is made possible by data-based training of the normal state and the resulting error for new data points. In conjunction with data preparation and consideration of the total error sum, the real anomalies can be detected, while irrelevant outliers are ignored. A further advantage is that the transitions between the training and application phases can be completely automated.Training continues until the model parameters converge, eliminating the need for manual pre-selection of good / bad data. During use, a continuous check is performed to determine whether the reference model is still up-to-date, and a new training process is automatically initiated. The expected relationship between the signals is trained using machine data, eliminating the need for detailed consideration of the converter parameters. A complex expert model is replaced by a simple data model. This means easy scalability and applicability to any assets, i.e., to different electrical machines. The features of the individual claimed or described objects and methods can be readily combined with one another. The invention is illustrated and explained in more detail below using figures as examples.The features shown in the figures can be combined by a person skilled in the art to form new embodiments without departing from the invention. They show: FIG. 1 the use of models and FIG. 2 training phases and application phases. 202214353 Foreign version 12 The illustration in FIG. 1 shows the use of models, in particular the interaction of automated training and application phases. The illustration in FIG. 2 shows training phases and application phases together with a transition time via a timeline 7. This timeline shows the training 1, 1' and the application of the reference model (RM) in timeline 7. After a training phase 1' of the reference model (model), this is finished and is used in application phase 2. Application phase 2 has normal phases 6, 6', 6'', which are unremarkable, and phases that are not normal, i.e. represent an anomaly.These phases with the anomaly are the anomaly phases 5, 5', 5''. Through at least one of the anomaly phases it can be recognized that the model (RM) is outdated. This results in a transition period 8 in which the model is outdated. After that, a new training 1' takes place, with a subsequent application 2' of the model with a new first normal phase 6 in the application 2'. The illustration in Figure 1 shows the sequence in a block diagram. In particular, a quantity of machine data is used to train a reference model (model) which describes a normal state in a training phase 1. The pre-processed, relevant data is used for this. Training continues until it is finished. For this purpose, the model is queried 3 iteratively, or in a loop, and an answer 4, 5 is determined. When the model is finished, an application phase 2 follows. In this iteratively, orin a loop, it is queried whether the model is old, i.e., outdated. This results in the answers 4' for yes and 5' for no, which lead to the corresponding subsequent loop. The answer as to whether an outdated model is present is determined from an evaluation of the detected anomalies. During operation, the trained model is used in application phase 2 to classify the time points as normal or abnormal. For this purpose, for example, an error resulting from the 202214353 Foreign Version 13 reference model is added to a total error sum and evaluated based on a fixed threshold. The current timestamp is classified as normal or abnormal. The interaction of the two phases is illustrated in Figure 1 and Figure 2. If data is available, especially from the past, a training phase can begin. As soon as the reference model converges, i.e.As soon as the model parameters no longer change significantly, it is considered fully trained and can be used for new data points in the application phase. During the application phase, the time points are classified as normal or abnormal. This way, unusual conditions are detected and reported to a user. If the unusual condition persists, it can be assumed that there has been a fundamental system or operational change. The originally trained "normal state" no longer reflects reality and must be retrained. A transition period (e.g., one week or one month) ensures that the change is made visible as an anomaly. A response can be made based on the procedure described. In response, the reference model can be updated independently, i.e., automatically, after which a new normal state is defined.

Claims

202214353 Foreign version 14 patent claims 1. Method for monitoring an electrical machine, wherein the monitoring is carried out for recurring operating states of the electrical machine, wherein speed-related data is used for monitoring, wherein a linear regression model is trained on the basis of training data, wherein a total error sum is calculated, wherein if a limit value is exceeded by the total error sum, an error is detected and / or a classification is carried out and / or a new model is trained. 2.Method according to claim 1, wherein monitoring of real data from a control of the dynamoelectric machine is carried out with expected data, wherein the data is speed-related, wherein deviations of the real data and the expected data are accumulated over time, wherein a classification of the behavior of the control is carried out using the deviations, wherein the classification of the control is monitored over time, wherein the expected data are changed depending on the monitoring of the classification.

3. Method according to claim 1 or 2, wherein the data relate to a speed and / or a torque of the electric machine.

4. Method according to one of claims 1 to 3, wherein the accumulation over time relates to minutes, hours, days and / or months.

5. Method according to one of claims 1 to 4, wherein a model relating to the electric machine is used to generate the expected data.Method according to claim 5, wherein a new model is generated depending on the classification. 202214353 Foreign version 15 7. Method according to claim 5 or 6, wherein the model and / or the new model are trained and / or are trained.

Images